Comma.ai Drive and Tour with George Hotz and Lex Fridman

00:00:00.000 | George Hotz gave me a demo of the Kama 2 on a ride back to the Kama AI offices in San Diego.

00:00:06.160 | Then we took a quick tour of the office. I found both the Kama 2 system and the entire Kama AI

00:00:14.640 | team and operation very impressive. Plus I got a 3D printer named after me. As one robot to another,

00:00:23.280 | we're in this together my brother. And now on to the ride and tour with George Hotz.

00:00:30.240 | I'm here with uh driving with George Hotz and uh looking at the Kama 2.

00:00:36.960 | Trying to find a stretch of road where we can enable it. Yeah just start looping back to the

00:00:43.760 | office. I mean I could enable it here but this isn't really where it shines. Yeah so we don't

00:00:49.600 | have we don't have our license stop signs yet. Um I don't want to ship it until we have a real uh

00:00:55.760 | holistic one. Until we have an end to end one. So you'll be able to then like stop at a stop sign.

00:01:01.440 | So the longitudinal control. Did you just enable it? I just enabled it yeah. So this is the uh so

00:01:07.280 | we're gonna have to stop at this light. But um how do you enable it? Uh cruise control. So that's

00:01:17.920 | that's the the integrated knobs on the on the car. Yeah we have a good we'll go on the 163 a bit and

00:01:25.840 | that's that's that's a good test of uh whether your level two system's any good. Supercruise

00:01:34.080 | actually can't do it. Supercruise can't do it because they're they're not good enough with roll

00:01:38.640 | on the road. Autopilot can do it and we can do it.

00:01:46.240 | How's San Diego in terms of quality of roads and lane markings and all that kind of stuff?

00:01:50.080 | Pretty mediocre like the rest of California. What's your favorite uh reason for using KAMA?

00:01:58.240 | Like when when do you find the most joy? Oh oh it's it's long trips. It's um you know

00:02:07.840 | driving from here to LA. You'll do it without a disengagement almost. You can you can you know

00:02:13.920 | it's two hours to LA from here. You could just go to LA. Uh get in your car and you just sit there

00:02:20.080 | and watch. And it's it's a it's a great time. All right so we can we can engage here. So you know

00:02:27.600 | watch. So it doesn't technically require perfect lane markings? No not at all. In fact you can do

00:02:37.840 | lane changes on roads that don't have lane markings because it's all end to end. So I have

00:02:43.120 | to get over three lanes. I'm just going to take control here to do it. Especially since there's

00:02:47.680 | a cop there. So let me just get over three lanes and then this road's a decent demo.

00:02:55.840 | Yeah so I think only us and Autopilot can do this road.

00:03:00.880 | Pretty curvy. Yeah and uh it's not just curvy it's also banked.

00:03:06.000 | Um and if you don't yeah if you don't do that well if you don't have a model that can detect

00:03:13.040 | that. Um so the longitudinal policy is actually just being done by the car.

00:03:18.320 | Um Hyundai has quite a competent ACC. Probably the best one. Uh so we haven't really felt the

00:03:24.560 | urge to move to our gas and brakes. There's a few cute things where which ours will do better

00:03:29.520 | and more holistically than than theirs. But um really it's about the turning.

00:03:38.000 | How hard is it to get it to be this smooth? Uh hard.

00:03:42.560 | Five years of engineering. All of this end to end huh?

00:03:47.760 | When we're detecting the lane lines they're used but they're more used as uh

00:03:53.440 | see like right there that's the stock system being too harsh on the brakes.

00:03:57.520 | If you're a human driving you would have been much more gradual on the brakes there.

00:04:04.880 | But we're really going to push to switch all the longitudinals to open pilot once we have

00:04:08.800 | end-to-end longitudinal and the same way we have end-to-end lateral.

00:04:11.120 | What's harder end-to-end longitudinal or end-to-end lateral?

00:04:16.240 | We put a lot more work into end-to-end lateral because there's a lot more subtlety in that.

00:04:19.680 | The car companies have managed to build competent longitudinal. We have competent

00:04:23.600 | longitudinal using old school policies. But lateral there's so much subtlety. For example

00:04:28.800 | like when a lane splits out should you follow that split? Should you stay in the center or

00:04:33.360 | should you stay to one side or so much subtlety there.

00:04:36.400 | And making it all smooth and not freak out. Yeah and this thing you know the reason I feel

00:04:45.280 | very confident keeping my hands off the wheel our torque limit is so much lower than Tesla's.

00:04:49.440 | So you can't can't do anything crazy? No we do injection tests where one person sits in that

00:04:56.160 | seat with a joystick and the other person keeps their hands off the wheel and that person you

00:05:00.320 | can't look at them they're free to jam the joystick in any direction and if you don't feel

00:05:04.160 | you know safe in that maneuver they take back over here because we're getting off here.

00:05:07.520 | But yeah that stretch of road is a is a is a competent demonstration of open pilot.

00:05:14.960 | Yeah that was impressive. It's not the high end. Maybe when we get to the city streets

00:05:19.200 | I'll engage on the city streets. Is city streets like a compelling use case or is that just kind of

00:05:25.760 | without it? Intuition builder. Without end-to-end longitudinal it's pretty like it's not really

00:05:32.320 | useful as product. So it's more just it's cool that it can do it and like we don't gate it off

00:05:38.640 | but most of open pilot is highways like most of the value real gain is highways.

00:05:44.080 | But also not just highways also like one lane each way kind of things like double yellow kind

00:05:53.840 | of roads. I was doing all like the mountains up in Colorado with this thing and just you get to look

00:05:58.800 | at the mountains before it drives. So like I can engage it here.

00:06:04.960 | Now remember the longitudinal policy is is Hyundai and it's not ideal for cities.

00:06:14.400 | But so again we're going to go through this intersection with no lines.

00:06:23.600 | See notice how there's no line there yet our placement is still almost human ideal.

00:06:28.000 | Yeah this is great. So there I mean there you see we did a stretch of

00:06:33.760 | you know a minute of city and perfectly human nothing sketchy.

00:06:37.280 | See places like that or where disengage on gas is really nice.

00:06:42.560 | Really easy to you know hand off you always know when that handoff is happening with the noise.

00:06:51.120 | It didn't like keep the ACC on or anything. God if it kept the ACC on and I make the turn and it's

00:06:55.840 | like trying to accelerate at a weird time you know just asking to have a bad time.

00:07:01.040 | I like the the sounds the visual it's like clear when it's on and when it's off.

00:07:11.760 | Love it. Well done. Damn that's impressive.

00:07:19.680 | That's really impressive. That's the intersections without thinking about it.

00:07:28.320 | Okay that was awesome.

00:07:32.480 | It's communicating the uncertainty with this movement a little bit.

00:07:40.240 | I like it. Humans actually wobble a little more than you think too.

00:07:43.920 | You notice it a lot more when they're this. Yeah that's an interesting point yeah. Yeah true.

00:07:50.000 | I think Jesse Levinson mentioned this on one of his interviews. He's like so you know we have all

00:07:58.800 | these we have the safety driver sit in the car and mark every tiny mistakes it made and then we also

00:08:03.760 | had the safety driver sit in a car that was being human driven and mark all the little tiny mistakes

00:08:07.920 | it made. So this is the comma office in downtown. The comma office is right here.

00:08:12.400 | We're at uh comma AI offices with uh Mr. George Harts. Can you give us a little tour? It doesn't

00:08:22.080 | light up anymore I don't know why. Do you think most of these dashboards are okay? They're almost

00:08:27.760 | at 30 million open pilot miles. Wow. Oh it's a hell of a number. So here here's

00:08:36.880 | dailies, weeklies and uh so actually yeah dailies are like maybe not 2,000 1.5.

00:08:42.320 | So those are used daily and then used weekly. Yeah. That's really cool. Weeklys are almost 2.5

00:08:52.080 | and monthly is a little over 3. This is how many devices are on our botnet in the last hour.

00:08:56.720 | Botnet last day, botnet last week. Is this all over the United States?

00:09:04.560 | This is no this is everywhere. So here's here's here's the maps. Here's the map of the world.

00:09:09.520 | Uh so we're everywhere in the world. So let's see. In the last 30 days we had 2,200 in the U.S.

00:09:22.400 | and then Korea, Canada, Taiwan, Japan. Japan's up there. This is our percent engaged.

00:09:33.680 | So percent of miles engaged is about 50 percent. Percent time engaged is about 30.

00:09:37.920 | Our disengagements are split pretty evenly between cancel, gas and break.

00:09:42.560 | Most of our segments are now coming from Comma 2s. Comma 2s. Comma 2 is the big the big breakout for

00:09:50.640 | this company. Yeah. Oh yeah. That's profitable. This is our live link to the Europe office.

00:09:55.920 | It comes as a multinational corporation. George just erased a bunch of top secret things.

00:10:02.960 | Censorship. The censors came in. Yeah. Look the mission's at the top of the board.

00:10:07.200 | Solve self-driving cars while delivering shippable intermediaries. Amen. What do we do badly? What

00:10:16.640 | can we do better? Right. Good questions. Open pilot 1.0. When is that coming up?

00:10:22.560 | When those things are done. PyTorch. Okay.

00:10:29.920 | I like how that's the bullet point. That's like move everything to PyTorch.

00:10:34.560 | Oh, sorry. It's not a data center. It's a compute cluster. For legal reasons,

00:10:40.000 | it's a compute cluster. It's a compute cluster. Okay. That sounds sexier too.

00:10:43.680 | Data ingestion, multi-chain. Cool. So this is like a vision board for OpenPilot 1.0.

00:10:51.600 | It's not going to happen for a while. But the Comma 2 is going to get OpenPilot 1.0.

00:10:56.000 | It's like what I said with the hardware. So this is all going to come to the Comma 2

00:10:59.360 | just with OTAs. OTA, over the air. We're going to get it. Yeah. I mean, we're getting close on a

00:11:04.960 | lot of these. I'm going to start lowering the wheel touch timeout. Alex, this is Alex. Hi.

00:11:11.280 | She's our COO. COO. Nice. Nice to meet you. She runs the back there stuff.

00:11:19.040 | What's the technical description of the back there stuff? You want to see? Yeah.

00:11:25.200 | I think it's more exciting than this stuff. It depends who you ask.

00:11:29.200 | We're ahead of all the secrets, right? Nick, do we have any secrets out?

00:11:35.040 | I'll blur it out. I'll blur the top secrets. This is production. Stress test. This is our

00:11:46.960 | final test. The Comma 2s hang out here for about 24 hours. We test them for their heat,

00:11:54.400 | their temperature, their light sensors. We do cosmetic checks, screen checks.

00:12:00.640 | Pretty exciting. You can see they all start over there as phones. Then they get retrofit.

00:12:08.160 | Then they go full house fit. You can see we have a class of 100 phones out right now.

00:12:18.080 | They're getting tested. It's kind of exciting actually. Lex, this is Mitchell. He's the

00:12:27.360 | latest addition to our machine learning team. That's Chris, also known as Virtually Chris.

00:12:38.800 | He's a famous YouTuber. Oh, wow. How did fame change you? I introduced you as a famous YouTuber.

00:12:47.840 | Yes, I was the open pilot YouTuber. Don't let it get to your head.

00:12:53.040 | That's our head of hardware. Awesome. What's that stage of the assembly?

00:13:04.960 | This is the final case. We have board testing. Then the retrofit phones get

00:13:12.400 | board put on them, soldered to them. We get the heat sink and the fan. Then we put the cases on

00:13:17.040 | them. Awesome. Yeah, they're taking cases off, screens off here. They're testing screens over

00:13:25.920 | here. Oven. Yeah, 3D printer. You want to show that? I will too. I'm afraid to take it in there.

00:13:32.640 | It's very hot and loud. It's also the home of the compute cluster.

00:13:39.760 | It's not a data center. It's not a data center. That's right. It's a compute cluster.

00:13:43.280 | Prepare yourself.

00:13:44.160 | It's the sound of productivity. Yeah. Oh my God, it's really hot. Fresh bag, cases.

00:13:59.120 | This is what you see on the outside of the Comma 2s.

00:14:07.040 | These are all of our printers.

00:14:08.080 | Yeah, I like how they have names.

00:14:14.240 | Meg. What's the purpose of the rings? They used to be for athletic activity, but now we have too

00:14:25.440 | many printers. You can't anymore. We just got about 20 printers in the last few months.

00:14:32.240 | They've come from all over the world. Literally, Poland, Germany, Peru, France.

00:14:38.960 | How long does it take to print a single one?

00:14:44.560 | Yes, we'll name number 49 for you, Lex. Thank you. We'll call him Lex. I told him he could be

00:14:53.440 | number 49. I could be. So great, 49's not working right now, but I promise 49 will work. It'll be

00:14:59.840 | Lex number 49. I appreciate that. That would mean a lot to me.

00:15:03.600 | This is cool. So unfortunately, because we got to keep the bugs and dust out, they're covered, but

00:15:14.000 | 1U. I already forgot their official name. The compute cluster. So these are our CPU machines.

00:15:26.400 | We have three racks that look like that. Two petabytes of spinning disks.

00:15:31.680 | We're at 3080s. Our 3080s are in only desktop computers for non-business purposes. These are

00:15:41.200 | our GPU machines. What's in there? V100s.

00:16:00.000 | It's like a farm. Every day we kind of go out.

00:16:01.920 | 21st century farm. We got to hire a good farm bill player to come farm us.

00:16:08.320 | These are the ones that are ready to ship.

00:16:12.720 | I think these are all of our growing harnesses. These are all the hardware support.

00:16:18.160 | Bosch, Deliver, Nynex, Subaru, Toyota, all the Hyundais. All the way to J,

00:16:27.520 | Nissan, VW. All these are assembled by hand?

00:16:34.080 | Yeah. That's cool.

00:16:35.600 | That's amazing. That's legit. That's awesome.

00:16:38.960 | Yeah, I don't know.

00:16:39.920 | That's the office.

00:16:42.880 | Have you been for a ride among the cars?

00:16:44.400 | Yeah. He took me for a ride. It was awesome. It was really surprising how well it worked.

00:16:48.800 | Where'd you take him on the highway?

00:16:50.400 | We did like the 163.

00:16:51.600 | In the city a little bit too. It held up.

00:16:55.040 | Definitely.

00:16:56.320 | You didn't show him any secrets, though, did you?

00:16:57.600 | No, I didn't show him any secrets. It means like keep on driving, you know?

00:17:01.920 | That's why we replaced it with a heart. We're all big Elon fans.

00:17:04.960 | Elon's success is our success.

00:17:10.080 | If iOS succeeds, Android succeeds.

00:17:15.040 | Thanks.

00:17:16.580 | you

00:17:17.620 | you

00:17:17.700 | you

00:17:18.200 | you

00:17:19.660 | you

00:17:20.700 | you

00:17:20.780 | you

00:17:21.280 | [BLANK_AUDIO]

00:17:26.360 | [BLANK_AUDIO]

Comma.ai Drive and Tour with George Hotz and Lex Fridman

Chapters